Physical activity plays an important role in controlling obesity and maintaining healthy living. It becomes increasingly important during a pandemic due to restrictions on outdoor activities. Tracking physical activities using miniature wearable sensors and state-of-the-art machine learning techniques can encourage healthy living and control obesity. This work focuses on introducing novel techniques to identify and log physical activities using machine learning techniques and wearable sensors. Physical activities performed in daily life are often unstructured and unplanned, and one activity or set of activities (sitting, standing) might be more frequent than others (walking, stairs up, stairs down). None of the existing activities classification systems have explored the impact of such class imbalance on the performance of machine learning classifiers. Therefore, the main aim of the study is to investigate the impact of class imbalance on the performance of machine learning classifiers and also to observe which classifier or set of classifiers is more sensitive to class imbalance than others. The study utilizes motion sensors' data of 30 participants, recorded while performing a variety of daily life activities. Different training splits are used to introduce class imbalance which reveals the performance of the selected state-of-the-art algorithms with various degrees of imbalance. The findings suggest that the class imbalance plays a significant role in the performance of the system, and the underrepresentation of physical activity during the training stage significantly impacts the performance of machine learning classifiers.
Keywords: digital health; e-health; machine learning; pandemic; performance evaluation; physical activity.